Jacques Féjoz, Université Paris-Dauphine & Observatoire de Paris, France,
[email protected]
Abstract: The theory of Kolmogorov, Arnold, and Moser (KAM) consists of a set of results regarding the persistence of quasiperiodic solutions, primarily in Hamiltonian systems. We bring forward a “twisted conjugacy” normal form, due to Herman, which contains all the (not so) hard analysis. We focus on the real analytic setting. A variety of KAM results follow, including most classical statements as well as more general ones. This strategy makes it simple to deal with various kinds of degeneracies and symmetries. As an example of application, we prove the existence of quasiperiodic motions in the spatial lunar three-body problem
KAM theory consists of results regarding the existence of quasiperiodic solutions, primarily in Hamiltonian systems. It was initiated by Kolmogorov in 1954, before being further developed by Arnold, Moser, and others.
The phase space of an integrable Hamiltonian system is foliated by Lagrangian invariant tori carrying a resonant or nonresonant quasiperiodic dynamics. Kolmogorov’s theorem asserts that, for any perturbation of the Hamiltonian, many nonresonant quasiperiodic Lagrangian invariant tori persist [56]. Kolmogorov’s proof consists in looking for a strongly nonresonant invariant torus and solving the corresponding functional equation using Newton’s algorithm in a (non-Banach) functional space of infinite dimension. Poincaré wrote that the convergence of perturbation series looked very unlikely. Others, among Weierstrass or Bogoliubov–Krylov, failed to prove it. Proving the convergence of perturbation series is difficult due to the accumulation, in this context, of “small denominators”. On the other hand, looking for invariant tori more geometrically as one would in the general theory of invariant manifolds, without prescribing a precise dynamics on the torus, fails severely because a resonant invariant torus does not persist under a generic perturbation (see [76] or [19, Section 2.6 and Chapter 11]). It was a stroke of genius of Kolmogorov both to imagine the correct statement and to realize that the Newton algorithm could be implemented in a family of Banach spaces and beat the effects of small denominators.
The invariant torus theorem has many applications in mathematical physics and mechanics. For example, Arnold’s theorem tells us that the (1 + n)-body planetary problem, where one body (standing for the Sun) has a (hugely) larger mass than the others (standing for planets), has many quasiperiodic solutions [2, 37]. In the circular restricted three-body problem, the regular limit where one of the masses vanishes while the other two describe circular motions around their center of mass, those invariant tori separate energy levels and, thus, confine all neighboring solutions, resulting in a resounding stability property. In the same line of thought, Herman has shown that Boltzmann’s ergodic hypothesis fails for a generic Hamiltonian, because codimension-1 invariant tori prevent energy levels to be ergodic [99]. Yet, invariant tori theorems seldom apply directly, due to symmetries and degeneracies, and to the strength of nonintegrability. Should KAM theory apply at all (see [15]), refined versions are usually needed.
Bibliographical comments. For background in Hamiltonian systems, excellent references are the books of Arnold [3], Arnold et al. [4], Guillemin and Sternberg [46], Knauf [55], Meyer and Hall [67], Siegel and Moser [92], or Sternberg [95].
There exist many surveys of KAM theory, among which we recommend those of Bost [11], Chierchia [21], Pöschel [78], or Sevryuk [88, 89, 91]. Dumas’s book [28] is an interesting, historical account of the subject. Several results emphasized in the present paper, including the twisted conjugacy Theorem 1.1, were already proved in [37] in the smooth setting.
Here are some examples of refinements or extensions, sometimes spectacular:
–quantitative versions [23]
–persistence of lower dimensional tori [30] even without controling the first order normal dynamics [13]
–persistence under degenerate torsion [32]
–global (nonperturbative) versions for diffeomorphisms of the circle [50] or cocycles [6]
–various kinds of linearizations of cocycles [5] or of interval exchange maps [65]
–the dynamics of generic Lagrangian invariant tori [99]
–weak KAM theory [34]
–reversible systems [90]
–non-Hamiltonian perturbations [66]
–Hamiltonian partial differential equations [9, 57]
(see references therein).
Let ℋbe the set of germs along T0 = ?n ×{0} of real analytic functions (Hamiltonians) in ?n × ℝn = {(θ, r)}. The Hamiltonian vector field associated with H ∈ ℋ is
For any given vector α ∈ ℝn, let ?(α) be the affine space of Hamiltonians K ∈ ℋ of the form
for some (nonfixed) c ∈ ℝ; O(r2) stands for the remainder (depending on θ) of the expansion in power of r. The space ?(α) consists exactly of Hamiltonians for which T0 is invariant (r|̇ r=0 = 0) and carries a linear flow with velocity α (θ̇|r=0 = α).24
Let ? be the set of germs along T0 of exact symplectic real analytic isomorphisms of the form25
where φ is an isomorphism of ?n fixing the origin and S is a function of ?n vanishing at the origin. The goal being to find invariant tori close to T0 and carrying a linear flow of frequency α, φ allows us to make changes of coordinates at will on the Lagrangian torus T0, while S allows us to bring back to the zero section any graph over T0, of 0-average and sufficiently close to the zero section.
In the next theorem, we assume that α is Diophantine:
for some fixed γ, τ > 0; we have set |k| = |k1|+⋅ ⋅ ⋅+|kn|. We will call Dγ,τ the set of such vectors. Dγ,τ is nonempty if and only if τ ≥ n − 1 (Dirichlet’s theorem) and, if τ > n − 1 and γ → 0, the complement of Dγ,τ within a ball has measure O(γ); hence, ∪γDγ,τ has full measure [81].
Theorem 1.1 (Herman). If Ko ∈ ?(α) and if H ∈ ℋ is close enough to Ko, there is a unique (K, G, β) ∈ ?(α) × ? ×ℝn such that
We will prove Theorem 1.1 in the next two sections. The statement calls for some remarks.
–The frequency being a conjugacy invariant of quasiperiodic flows, the counterterm β ⋅ r, which allows us to tune the frequency, is necessary. Yet it breaks the dynamical conjugacy between K and H and does not comply H with having an invariant torus, as K does.We call this normal form a twisted conjugacy. The geometrical contents of the theorem are that locally the set of Hamiltonians possessing anα-quasiperiodic torus is a submanifold of finite codimension if α is Diophantine (it has infinite codimension if α is not). The counter-term is the finite-dimensional obstruction to conjugacy to a Hamiltonian of ?(α) and can be imaged as a simple control to preserve a torus of the same frequency and cohomology class as that of Ko.
–In general, one cannot expect H to be of the form
this would show that having a Diophantine invariant torus is an open property,which is wrong, as the following example shows.
Consider the Hamiltonian H = α ⋅ r, α ∈ ℝ2. All the tori r = cst are invariant. Bythe first arbitrarily small perturbation, we may assume that α is resonant: k ⋅α = 0for some k ∈ ℤ2 {0}. Then, add a resonant monomial:
H = α ⋅ r − ϵ sin(2πk ⋅ θ) .
The vector field is
So, the solution through (0, r) at time t = 0 is
t → (tα, r + 2πϵtk) .
So, if ϵ > 0, this solution is unbounded and prevents any invariant torus (among graphs over T0) to exist.
Exercise 1.2. Deduce Arnold’s normal form for vector fields υ on the torus ?n close to a Diophantine rotation [2, 11], from the twisted conjugacy theorem. Hint: Apply the twisted conjugacy theorem to the Hamiltonian H (θ, r) = r ⋅ υ on ?n ×ℝn. Then, check that the zero section is invariant by the corresponding isomorphism G, using an argument of Lagrangian intersection [37].
Bibliographical comments.
–Computing the codimension of a group orbit is sometimes imprecisely called the “method of parameters.” It is commonplace in singularity theory. In KAM theory, where the codimension often turns out finite, it has been fruitfully used in a number of works, among which Arnold’s normal form of vector fields on the torus (the paradigmatic, founding example) [2, 70, 99], Moser’s normal form of vector fields [71] (which encompasses many natural subcases [14, 66, 97] but which has been much overlooked for 30 years), Chenciner’s work on bifurcations of elliptic fixed points [16–18] or Eliasson–Fayad–Krikorian’s study of the neighborhood of invariant tori [33]. The method of parameters allows us to first prove a normal form theorem which does not depend on any nondegeneracy assumption, but which contains all the hard analysis; the remaining, finite-dimensional problem is then to show that the frequency offset vanishes, using a nondegeneracy hypothesis. This last step was probably not well understood before the late 1980s [30, 60, 61, 84, 88]. The method fails for other kinds of dynamics than the quasiperiodic one on the torus because generically there are infinitely many new obstructions (the right-hand side of the cohomological equation should have 0-average on periodic orbits) at each step of the Newton algorithm [43].
–The normal form of Theorem 1.1, which was advertised by Herman in the 1990s (M. Herman seemingly did not know Moser’s normal form), in particular in his lectures on Arnold’s theorem at the Dynamical System Seminar in Université Paris VII, can be seen as a particular case of Moser’s normal form, when the vector field is Hamiltonian, then giving more precise information [38, 66]. A proof in the smooth category can be found in [37]. The lesser rigidity there allows us not to introduce deformed norms.
Let
We want to solve the following equation between Hamiltonians:
for H close to ϕ(Ko , id, 0) = Ko. The twisted conjugacy theorem, thus, reduces to prove that ϕ is invertible, keeping in mind that
–ifϕ is formally defined on the whole space ?(α) × ? × ℝn, it is only if G is close enough to the identity, with respect to the width of analyticity of K, that ϕ(K, G, β) is analytic on a neighborhood of T0,
–Equation (2.1) is really of interest to us only if it holds on a neighborhood of G−1(T0), a domain depending on the unknown G.
Note that ℋ and ℝn are trivially vector spaces, while
–?(α) is an affine space, directed by the vector space ℝ+ O(r2)
–and?, while being a groupoid with semidirect product law given by G2 ∘ G1 = will rather be identified, locally in a neighborhood of the identity, to an open set of the affine space passing through the identity and directed by the linear space {(φ − id, S)}, where υ = φ − id: ?n → ℝn and S : ?n → ℝ are analytic and vanish at the origin.
We will invert ϕ using the Newton algorithm, which consists in iterating the operator
Each step of the induction requires to invert the linearized operator ϕ′(x), not only at x0 = (Ko, id, 0), but at some unknown x in the neighborhood of x0, that is, to solve the linearized equation
where δH is the data, (K, G, β) is a parameter, and the unknowns are the “tangent vectors” δK ∈ ℝ⊕ O(r2), δG (geometrically, a vector field along G) and δβ ∈ ℝn. Precomposing with G−1 modifies the equation into an equation between germs along the standard torus T0 (as opposed to the G-dependant torus G−1(T0)):
where we have set G ̇ = δG∘G−1 (geometrically, a germ along T0 of tangent vector field) and H ̇ = δH ∘ G−1. It is a key point in measuring norms that we are interested in the neighborhood of T0 on one side of the conjugacy, and in the neighborhood of G−1(T0) on the other side (Figure 1).
Using the additional notations (in which ∗≥k stands for a function in O(rk ), depending on θ):
and the fact that
and identifying the Taylor coefficients in equation (2.4) yield the following three equations:
The first equation aims at infinitesimally straightening the would-be invariant torusof Ḣ 0, the second equation at straightening its dynamics, and the third at equating higher order terms. Due to the symplectic constraint, the first two equations are coupled (the Hamiltonian lift of the vector field φ̇ having a nontrivial component in the r-direction), whereas in the context of general vector fields, the system is triangular. We will now show the existence of a unique solution to this system of equations, and derive estimates of the solution, within some appropriate functional setting.
Let
be the complex extension of ?n of “width” s, and
for functions f which are real holomorphic on the interior of and continuous on ; such functions form a space which is Banach (there are other possible choices here, e.g., one could consider the space of functions which are real holomorphic on ). We extend this definition to vector-valued functions by taking the maximum of the norms of the components (and, consistently, the ℓ1-norm for “dual” integer vectors, e.g., k ∈ ℤn). Similarly, let be a complex neighborhood of the origin in ℝn of “width” s:
We will call the Banach space of functions which are continuous on and real holomorphic on the interior.
Let
–ℋs = ℋ ∩ ? (endowed with the supremum norm | ⋅ |s)
–?s(α) = ?(α) ∩ ? ⊂ ℋs.
– be the subset of ? consisting of isomorphisms sG σ≃ (φ, S) such that φ − id ∈ and S ∈ and G is σ-polynomially-close to the identity, that is,
for some fixed CG > 0 and kG > 0 to be determined later.
Lemma 2.1 (Linearized equation). If x is close enough to x0, equation (2.4) possesses a unique solution x ̇ = (δK, Ġ , δβ). Moreover, there exist C′ , τ′ > 0 such that, for all s, σ,
where Cˊ depends only on n, τ, provided K, G−1 and β are bounded on
Proof. First assume that δβ ∈ ℝn is given with |δβ ⋅ r ∘ G−1| ≤ Cst |Ḣ |s+σ, and replace equation (2.6) by
where δβ̂ ∈ ℝn is an additional unknown; as elsewhere in this proof, Cst stands for a constant, to which we do not want to give a consistent name, and which depends only on n, τ, and |(K − α ⋅ r, G−1 − id, β)|s+σ.
–Averagingequation(2.5)yieldsδc = ∫?n (Ḣ 0 + Sˊ ∘ φ−1 ⋅ δβ) dθ; hence,
–According to Lemma B.1, equation (2.5) has a unique solution δS̃ having 0-average, with
Then, the unique solution vanishing at the origin, δS = δS̃ − δS̃(0), satisfies the same estimate (up to an unessential factor 2 which we absorb in the constant).
Note that the estimates hold for all s, σ (at the expense of possibly having an infinite right-hand side). We now proceed similarly with equation (2.9):
–The average yields δβ̂ = ∫?n (Ḣ 1 − 2Ṡˊ ⋅ Q − φˊ ∘ φ−1 ⋅ δβ); hence, using Cauchy’s inequality,26
–The average-free part determines δφ with
Using Cauchy’s inequality and the fact that Qo is given, we see that
where as before the constant depends only on n, τ, and |Qo|s+σ.
Equation (2.7) can then be solved explicitly:
and whence
We have built a map δβ ↦ δβ̂ in the neighborhood of δβ = 0. It is affine and, when φ is close to the identity, invertible. Thus, there exists a unique δβ such that δβ̂ = 0, which satisfies
The claim follows, with τ′ = τc(τc + 1) + 1 and C′ = Cst/γ2 for some constant Cst independent of γ.27
The lemma may be rephrased: the linear operator ϕ′(x) has a unique local inverse ϕ′(x)−1, with the given estimates.
Let
Taylor’s formula says that
where we have set xt = x + t δx (0 ≤ t ≤ 1); hence,
Lemma 2.2 (Remainder). If |Ġ |s+σ ≤ σ/2,
Proof. Let δ2ϕ = ϕ''(K, G, β) ⋅ (δK, δG, δβ)2. We have
hence,
so
note here that δ2ϕ is computed in (K, G, β), and it is then precomposed by G−1.
Now, if xt = (Kt , Gt , βt),
Since
whence the wanted estimate, using (2.10).
It remains to show that the iterated images
of the Newton map (2.2) are defined for n ∈ ℕand converge to some (K, G, β) ∈ ?(α)× ?×ℝn such that H = K ∘ G + β ⋅ r in the neighborhood of G−1(T0), provided H is close enough to Ko. Namely, we will assume that Ko ∈ ?s+σ(α), H ∈ ℋs+σ for some fixed s, σ with 0 < s < s + σ ≤ 1, and
for some ϵ > 0. This is the goal of the next section.
We first give an abstraction of our problem and will afterward show how it allows us to complete the proof of the twisted conjugacy theorem.
Let E = (Es)0<s<1be a decreasing family of Banach spaces with increasing norms |⋅ |s, and = {x ∈ Es , |x|s< ϵ}, ϵ > 0, be its balls centered at 0. Let (Fs) be ananalogous family, and ϕ: → Fs, s < s + σ, ϕ(0) = 0, be maps of Eclass C2, commuting with inclusions.
On account of composition operators, we will assume that there are additional, deformed norms |⋅|x,s, x ∈ Int 0 < s < 1, satisfying
and we will phrase our hypotheses on ϕ in terms of these norms.
Define
Assume that, if x ∈ the derivative Eϕ′(x) : Es+σ → F s has a right inverseϕ′(x)−1 : Fs+σ → Es, and
with C′ , C″, τ′ , τ″ ≥ 1. Let C := C′C″ and τ := τ′ + τ″.
The important fact in the Newton algorithm below is that the index loss σ can be chosen arbitrarily small, without s itself being small, provided the deformed norm substitutes for the initial norm of the spaces F s. The initial norm | ⋅ | s of F s is here only for the practical purpose of having a fixed target space, to which perturbations belong.
Theorem 3.1. ϕ is locally surjective and, more precisely, for any s, η, and σ, with η < s,
In other words, ϕ has a right inverse
Some numbers s, η, and σ, and being given, let
Proof of the theorem. Now, let s, η, and σ be fixed, with η < s and for someϵ.We will see that if ϵ is small enough, the sequence x0 = 0, xn := f n(0) is defined forall n ≥ 0 and converges toward some preimage of y by ϕ.
Let (σn)n≥0 be a sequence of positive real numbers such that 3 Σ σn = σ, and (sn)n≥0 be the sequence decreasing from s0 := s + σ to s defined by induction by the formula sn+1 = sn − 3σn.
Assuming the existence of x0, . . . , xn+1, we see that ϕ(xk) = y + Q(xk−1, xk); hence,
Further assuming that |xk+1 − xk|sk ≤ σk, the estimate of the right inverse and Lemma 10.2 entails that
The estimate
and the fact, to be checked later, that ck ≥ 1 for all k ≥ 0, show
Since Σn≥0 ρ2n ≤ 2ρ if 2ρ ≤ 1, and using the definition of constants ck’s, we get a sufficient condition to have all xn’s defined and to have Σ|xn+1 − xn|s ≤ η:
Maximizing the upper bound of ϵ under the constraint 3 Σn≥0 σn = σ yields σk := A posteriori it is straightforward that |xn+1 − xn|sn ≤ σn (as earlier assumed to apply Lemma 10.2) and cn ≥ 1 for all n ≥ 0. Besides, using that Σ k2−k = Σ2−k = 2, we get
whence the theorem.
Remark. The two competing small parameters η and σ being fixed, our choice of the sequence (σn) maximizes ϵ for the Newton algorithm. It does not modify the sequence (xk) but only the information we retain from (xk).
Exercise 3.2 (End of proof of Theorem 1.1). Complete the proof by checking that
–A similar statement as Theorem 3.1 holds if ϕ is defined only on a ball of polynomial radius with respect to the width of analyticity (see (2.8)).
– and βn are bounded along the induction (in order to justify the repeated use of estimates of Lemmata 2.1 and 2.2,which are not uniform as assumed in (3.1)). Hint: Use the fact that
the estimate of G ̇ in the induction and the estimate of Proposition B.2 in appendix B.
Corollary 3.3. The size of the allowed perturbation is polynomial in the Diophantine constant γ (see (1.1)).
Exercise 3.4. What is the domain of ψ in F S? Hint: Optimize the function ϵ(η, σ) under the constraint s + σ = S.
Bibliographical comments.
–The seeming detour through Herman’s normal form reduces Kolmogorov’s theorem to a functionally well-posed inversion problem, as opposed to Zehnder’s (remarkable) work [100, 101]. One may compare the present stragegy and Zehnder’s in the following way. Inverting the operator
(see equation (2.1)) is equivalent to solving the implicit function
But ϕ happens to be a local diffeomorphism, while ∂F/∂(K, G, β) is invertible in no neighborhood of (Ko , id, 0). This is why Zehnder had to deal with approximate inverses. The drawback of focusing on the equation ϕ(K, G, β) = H is that we need it to be satisfied on a domain which depends on G.
As Zehnder, we have encapsulated the Newton algorithm in an abstract inverse function theorem, à la Nash–Moser theorem. The algorithm indeed converges without very specific hypotheses on the internal structure of the variables (see Exercise 3.2, though). At the expense of some optimality, ignoring this structure allows for simple estimates and control of the bounds, and for solving a whole class of analogous problems with the same toolbox (lower dimensional tori, codimension-1 tori, Siegel problem, as well as some problems in singularity theory).
–The fast convergence of the Newton algorithm makes it possible to beat the effect of small denominators and other sources of loss of width of analyticity. It has proved unreasonably efficient compared to other lines of proof in KAM theory such as direct proofs of convergence of perturbation series [31] or proofs via renormalization [27]. Another approach relies on the method of periodic approximation and on simultaneous Diophantine approximations [12]. Still another alternative to Newtons’s algorithm consists, at each step of the induction, in solving a (nonlinear) finite-dimensional approximation of the functional equation (2.1) using Ekeland’s variational principle [29].
–The arithmetic condition is not optimal. Indeed, solving the exact cohomological equation at each step is inefficient because the small denominators appearing with intermediate-order harmonics deteriorate the estimates, whereas some of these harmonics could have a smaller amplitude than the error terms and thus would better not be taken care of. Even stronger, Rüssmann and Pöschel have noticed that, at each step, it is worth neglecting part of the low-order harmonics themselves (to some carefully chosen extent). Then the expense, a worse error term, turns out to be cheaper than that the gain, namely, the right-hand side of the cohomological equation now has a smaller size over a larger complex extension. This makes it possible, with a slowly converging sequence of approximations, to show the persistence of invariant tori under some arithmetic condition which, in one dimension, is equivalent to Brjuno’s condition [79].
–The analytic (or Gevrey) category is simpler than Hölder or Sobolev categories, in Nash–Moser theory, because the Newton algorithm can be carried out without intercalating smoothing operators (cf. [11, 48, 68, 87]). On the other hand, the analytic category is more complicated because of the absence of cut-off functions, which forces us to pay attention to the domain of definition of the Hamiltonian more carefully (cf. [37]).
–The method of Jacobowitz [52] (see [69] also) in order to deduce an inverse function theorem in the smooth category from its analog in the analytic category does not work directly, here. The idea would be to use Jackson’s theorem, in approximation theory to characterize the Hölder spaces by their approximation properties in terms of analytic functions and, then, to find a smooth preimage x by ϕ of a smooth function y as the limit of analytic preimages xj of analytic approximations yj of y. However, in our inverse function theorem we require the operator ϕ to be defined only on balls σBs+σ with shrinking radii when s + σ tends to 0. This domain is too small in general to include the analytic approximations yj of a smooth y. Such a restriction is inherent in the presence of composition operators. The problem of isometric embeddings is simpler, from this viewpoint.
In the proof of Theorem 3.1, we have built right inverses commuting with inclusions. The proof shows that ψ is continuous at 0; due to the invariance of the hypotheses of the theorem by small translations, ψ is locally continuous.
We further make the following two assumptions:
–Themapsϕ′(x)−1 : Fs+σ → Es are left (as well as right) inverses (in Theorem 1.1 we have restricted to an adequate class of symplectomorphisms).
–The scale(| ⋅ |s) of norms of (Es) satisfies some interpolation inequality:
(according to the sentence after the statement of Corollary C.2 in Appendix C, thisestimate is satisfied in the case it interests to us, since σ + log(1 + σ/s) ≤ σ̃).
Lemma 4.1 (Lipschitz regularity). If σ < s and y, with ϵ = 2−14τC−3σ3τ,
In particular, ψ is the unique local right inverse of ϕ, that is, it is also the local left inverseof ϕ.
Proof. Fix η < ζ < σ < s; the impatient reader can readily look at the end of the proof how to choose the auxiliary parameters η and ζ more precisely.
Let ϵ = 2−8τC−2ζ 2τη, and y, According to Theorem s3.1F, x := ψ(y) and x ̂ := ψ( ŷ ) are in provided the condition, to be checked later, that η < s+ σ−ζ . In particular, we will use a priori that
We have
and, according to the assumed estimate on ϕ′(x)−1 and to Lemma 10.2,
In the norm index of the last term, we will coarsely bound |x̂ − x|s by 2η. Additionally,using the interpolation inequality:
yields
Now, we want to choose η small enough so that
–First, σ≤ σ − ζ , which implies |x̂ − x|s+σ̃≤ 2η. By definition of σ̃, it suffices to have
–Second,2−1Cζ−τ 2η ≤ 1/2, or which implies that 2−1Cζ −τ|x̂ −x|s+σ̃ ≤ 1/2, and hence |x̂ − x|s ≤ 2C′σ−τ′ |y− y|x,s+σ.
A choice is whence the value of ϵ in the statement.
Proposition 4.2 (Smoothness). For every σ < s, there exists ϵ, C1 such that for every
Moreover, the map defined locally by ψ′(y) = ϕ′(ψ(y))−1 iscontinuous and, if for all σ, so is
Proof. Fix ϵ as in the previous proof and y, Let x = ψ(y), η = y ̂ − y, ξ = ψ(y + η) − ψ(y) (thus, η = ϕ(x + ξ) − ϕ(x)), and Δ := ψ(y + η) − ψ(y) − ϕ′(x)−1η.
Definitions yield
Using the estimates on ϕ′(x)−1 and Q and the latter lemma,
for some σ′ tending to 0 when σ itself tends to 0, and for some C1 > 0 depending on σ. Up to the substitution of σ by σ′, the estimate is proved.
The inversion of linear operators between Banach spaces being analytic, y ↦ ϕ′(ψ(y))−1 has the same degree of smoothness as ϕ′.
Corollary 4.3. If π ∈ L(Es , V) is a family of linear maps, commuting with inclusions,into a fixed Banach space V, then π ∘ ψ is C1 and (π ∘ ψ)′ = π ⋅ (ϕ′ ∘ ψ)−1.
This corollary is used with π : (K, G, β) → β in the proof of Theorem 1.1.
It will later be convenient to extend ϕ−1 to non-Diophantine vectors α. Whitneysmoothness is a criterion for such an extension to exist [94, 98].
Suppose ϕ(x) = ϕα(x) now depends on some parameter α ∈ Bκ (the unit ball ofℝκ),
–that the estimates assumed up to now are uniform with respect to α over some closed subset D ⊂ ℝκ,
–andthatϕ is C1 with respect to α.
We will denote ψα the parametrized version of the inverse of ϕα.
Proposition 4.4 (Whitney-smoothness). If s, σ, and ϵ are chosen like in Proposition 4.2,the map is C1-Whitney-smooth and extends to a map of class C1. If ϕ is Ck, 1 ≤ k ≤ ∞, with respect to α, this extension is Ck.
Proof. Let If α, α + β ∈ D, xα = ψα(y) and xα+β = ψα+β(y), we have
Since ŷ ↦ ψα+β(ŷ) is Lipschitz (Lemma 4.1),
and, since α̂ ↦ ϕα̂ (xα) itself is Lipschitz, so is α ↦ xα.
Moreover, the formal derivative of α ↦ xα is
Expanding y = ϕα+β(xα+β) at β = 0 and using the same estimates as above show that
when β → 0, locally uniformly with respect to α. Hence, α ↦ xα is C1-Whitneysmooth, and, similarly, Ck-Whitney-smooth if α ↦ ϕα is.
Thus, by the Whitney extension theorem, the claimed extension exists. Note that Whitney’s original theorem needs two straightforward generalizations to be applied here: ψ takes values in a Banach space, instead of ℝ or a finite-dimensional vector space (see [44]); and ψ is defined on a Banach space, but the extension directions are of finite dimension.
Exercise 4.5 (Quasiperiodic time-dependent perturbations). Let ν ∈ ℝm be fixed. Consider the subspace ℋν of ℋ (in dimension 2(n + m)) consisting of Hamiltonians in
of the form
where H does not depend on Ψ. Since the corresponding Hamiltonian vector field has the component
ℋν may be imaged as the space of Hamiltonians on ?n × ℝn with quasiperiodic time dependence. Show that, if Ĥ ∈ ℋν and Ĥ = ϕ(K, G, (β, β′)) (with β ∈ ℝn and β′ ∈ ℝm), then
Further question: develop the KAM theory below in this particular case.
Exercise 4.6 (Control and persistence of tori). If H is close to an integrable Hamiltonian Ko = Ko(r), show that there is a smooth integrable Hamiltonian β = β(r) for every R such that T0 is a (γ, τ)-Diophantine invariant torus of Ko, H − β(r) ⋅ r has an invariant torus carrying a quasiperiodic dynamics with the same frequency.